Deep neural network-based strategy for optimal sensor placement in data assimilation of turbulent flow
Zhiwen Deng, Chuangxin He, Yingzheng Liu
Abstract
This paper focuses on the optimal sensor placement (OSP) strategy based on a deep neural network (DNN) for turbulent flow recovery within the data assimilation framework of the ensemble Kalman filter (EnKF). The assimilated flow field can be obtained using EnKF by optimizing the Reynolds-averaged Navier–Stokes (RANS) model constants. A feature importance layer was designed and used in a DNN to obtain the spatial sensitivity with respect to the RANS model constants. Two flow configurations experimentally measured using particle image velocimetry—i.e., a free round jet flow at Rej=6000 and a separated and reattached flow around a blunt plate at Reb=15 800—were selected as the benchmarks to demonstrate the effectiveness and robustness of the proposed strategy. The results indicated that the RANS models with EnKF augmentation were substantially improved over their original counterparts. A comprehensive investigation demonstrated that the selection of the five most sensitive sensors by DNN-based OSP can efficiently reduce the number of sensors and achieve a similar or better-assimilated performance over that obtained using all data in the entire flow field as observations.